xref: /aosp_15_r20/external/mesa3d/src/gallium/drivers/llvmpipe/lp_bld_depth.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /**************************************************************************
2  *
3  * Copyright 2009-2010 VMware, Inc.
4  * All Rights Reserved.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the
8  * "Software"), to deal in the Software without restriction, including
9  * without limitation the rights to use, copy, modify, merge, publish,
10  * distribute, sub license, and/or sell copies of the Software, and to
11  * permit persons to whom the Software is furnished to do so, subject to
12  * the following conditions:
13  *
14  * The above copyright notice and this permission notice (including the
15  * next paragraph) shall be included in all copies or substantial portions
16  * of the Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25  *
26  **************************************************************************/
27 
28 /**
29  * @file
30  * Depth/stencil testing to LLVM IR translation.
31  *
32  * To be done accurately/efficiently the depth/stencil test must be done with
33  * the same type/format of the depth/stencil buffer, which implies massaging
34  * the incoming depths to fit into place. Using a more straightforward
35  * type/format for depth/stencil values internally and only convert when
36  * flushing would avoid this, but it would most likely result in depth fighting
37  * artifacts.
38  *
39  * Since we're using linear layout for everything, but we need to deal with
40  * 2x2 quads, we need to load/store multiple values and swizzle them into
41  * place (we could avoid this by doing depth/stencil testing in linear format,
42  * which would be easy for late depth/stencil test as we could do that after
43  * the fragment shader loop just as we do for color buffers, but more tricky
44  * for early depth test as we'd need both masks and interpolated depth in
45  * linear format).
46  *
47  *
48  * @author Jose Fonseca <[email protected]>
49  * @author Brian Paul <[email protected]>
50  */
51 
52 #include "pipe/p_state.h"
53 #include "util/format/u_format.h"
54 #include "util/u_cpu_detect.h"
55 
56 #include "gallivm/lp_bld_type.h"
57 #include "gallivm/lp_bld_arit.h"
58 #include "gallivm/lp_bld_bitarit.h"
59 #include "gallivm/lp_bld_const.h"
60 #include "gallivm/lp_bld_conv.h"
61 #include "gallivm/lp_bld_logic.h"
62 #include "gallivm/lp_bld_flow.h"
63 #include "gallivm/lp_bld_intr.h"
64 #include "gallivm/lp_bld_debug.h"
65 #include "gallivm/lp_bld_swizzle.h"
66 #include "gallivm/lp_bld_pack.h"
67 
68 #include "lp_bld_depth.h"
69 #include "lp_state_fs.h"
70 
71 
72 /** Used to select fields from pipe_stencil_state */
73 enum stencil_op {
74    S_FAIL_OP,
75    Z_FAIL_OP,
76    Z_PASS_OP
77 };
78 
79 
80 
81 /**
82  * Do the stencil test comparison (compare FB stencil values against ref value).
83  * This will be used twice when generating two-sided stencil code.
84  * \param stencil  the front/back stencil state
85  * \param stencilRef  the stencil reference value, replicated as a vector
86  * \param stencilVals  vector of stencil values from framebuffer
87  * \return vector mask of pass/fail values (~0 or 0)
88  */
89 static LLVMValueRef
lp_build_stencil_test_single(struct lp_build_context * bld,const struct pipe_stencil_state * stencil,LLVMValueRef stencilRef,LLVMValueRef stencilVals)90 lp_build_stencil_test_single(struct lp_build_context *bld,
91                              const struct pipe_stencil_state *stencil,
92                              LLVMValueRef stencilRef,
93                              LLVMValueRef stencilVals)
94 {
95    LLVMBuilderRef builder = bld->gallivm->builder;
96    const unsigned stencilMax = 255; /* XXX fix */
97    struct lp_type type = bld->type;
98 
99    /*
100     * SSE2 has intrinsics for signed comparisons, but not unsigned ones. Values
101     * are between 0..255 so ensure we generate the fastest comparisons for
102     * wider elements.
103     */
104    if (type.width <= 8) {
105       assert(!type.sign);
106    } else {
107       assert(type.sign);
108    }
109 
110    assert(stencil->enabled);
111 
112    if (stencil->valuemask != stencilMax) {
113       /* compute stencilRef = stencilRef & valuemask */
114       LLVMValueRef valuemask = lp_build_const_int_vec(bld->gallivm, type, stencil->valuemask);
115       stencilRef = LLVMBuildAnd(builder, stencilRef, valuemask, "");
116       /* compute stencilVals = stencilVals & valuemask */
117       stencilVals = LLVMBuildAnd(builder, stencilVals, valuemask, "");
118    }
119 
120    LLVMValueRef res = lp_build_cmp(bld, stencil->func,
121                                    stencilRef, stencilVals);
122    return res;
123 }
124 
125 
126 /**
127  * Do the one or two-sided stencil test comparison.
128  * \sa lp_build_stencil_test_single
129  * \param front_facing  an integer vector mask, indicating front (~0) or back
130  *                      (0) facing polygon. If NULL, assume front-facing.
131  */
132 static LLVMValueRef
lp_build_stencil_test(struct lp_build_context * bld,const struct pipe_stencil_state stencil[2],LLVMValueRef stencilRefs[2],LLVMValueRef stencilVals,LLVMValueRef front_facing)133 lp_build_stencil_test(struct lp_build_context *bld,
134                       const struct pipe_stencil_state stencil[2],
135                       LLVMValueRef stencilRefs[2],
136                       LLVMValueRef stencilVals,
137                       LLVMValueRef front_facing)
138 {
139    LLVMValueRef res;
140 
141    assert(stencil[0].enabled);
142 
143    /* do front face test */
144    res = lp_build_stencil_test_single(bld, &stencil[0],
145                                       stencilRefs[0], stencilVals);
146 
147    if (stencil[1].enabled && front_facing != NULL) {
148       /* do back face test */
149       LLVMValueRef back_res;
150 
151       back_res = lp_build_stencil_test_single(bld, &stencil[1],
152                                               stencilRefs[1], stencilVals);
153 
154       res = lp_build_select(bld, front_facing, res, back_res);
155    }
156 
157    return res;
158 }
159 
160 
161 /**
162  * Apply the stencil operator (add/sub/keep/etc) to the given vector
163  * of stencil values.
164  * \return  new stencil values vector
165  */
166 static LLVMValueRef
lp_build_stencil_op_single(struct lp_build_context * bld,const struct pipe_stencil_state * stencil,enum stencil_op op,LLVMValueRef stencilRef,LLVMValueRef stencilVals)167 lp_build_stencil_op_single(struct lp_build_context *bld,
168                            const struct pipe_stencil_state *stencil,
169                            enum stencil_op op,
170                            LLVMValueRef stencilRef,
171                            LLVMValueRef stencilVals)
172 
173 {
174    LLVMBuilderRef builder = bld->gallivm->builder;
175    struct lp_type type = bld->type;
176    LLVMValueRef max = lp_build_const_int_vec(bld->gallivm, type, 0xff);
177 
178    assert(type.sign);
179 
180    unsigned stencil_op;
181    switch (op) {
182    case S_FAIL_OP:
183       stencil_op = stencil->fail_op;
184       break;
185    case Z_FAIL_OP:
186       stencil_op = stencil->zfail_op;
187       break;
188    case Z_PASS_OP:
189       stencil_op = stencil->zpass_op;
190       break;
191    default:
192       assert(0 && "Invalid stencil_op mode");
193       stencil_op = PIPE_STENCIL_OP_KEEP;
194    }
195 
196    LLVMValueRef res;
197    switch (stencil_op) {
198    case PIPE_STENCIL_OP_KEEP:
199       res = stencilVals;
200       /* we can return early for this case */
201       return res;
202    case PIPE_STENCIL_OP_ZERO:
203       res = bld->zero;
204       break;
205    case PIPE_STENCIL_OP_REPLACE:
206       res = stencilRef;
207       break;
208    case PIPE_STENCIL_OP_INCR:
209       res = lp_build_add(bld, stencilVals, bld->one);
210       res = lp_build_min(bld, res, max);
211       break;
212    case PIPE_STENCIL_OP_DECR:
213       res = lp_build_sub(bld, stencilVals, bld->one);
214       res = lp_build_max(bld, res, bld->zero);
215       break;
216    case PIPE_STENCIL_OP_INCR_WRAP:
217       res = lp_build_add(bld, stencilVals, bld->one);
218       res = LLVMBuildAnd(builder, res, max, "");
219       break;
220    case PIPE_STENCIL_OP_DECR_WRAP:
221       res = lp_build_sub(bld, stencilVals, bld->one);
222       res = LLVMBuildAnd(builder, res, max, "");
223       break;
224    case PIPE_STENCIL_OP_INVERT:
225       res = LLVMBuildNot(builder, stencilVals, "");
226       res = LLVMBuildAnd(builder, res, max, "");
227       break;
228    default:
229       assert(0 && "bad stencil op mode");
230       res = bld->undef;
231    }
232 
233    return res;
234 }
235 
236 
237 /**
238  * Do the one or two-sided stencil test op/update.
239  */
240 static LLVMValueRef
lp_build_stencil_op(struct lp_build_context * bld,const struct pipe_stencil_state stencil[2],enum stencil_op op,LLVMValueRef stencilRefs[2],LLVMValueRef stencilVals,LLVMValueRef mask,LLVMValueRef front_facing)241 lp_build_stencil_op(struct lp_build_context *bld,
242                     const struct pipe_stencil_state stencil[2],
243                     enum stencil_op op,
244                     LLVMValueRef stencilRefs[2],
245                     LLVMValueRef stencilVals,
246                     LLVMValueRef mask,
247                     LLVMValueRef front_facing)
248 
249 {
250    LLVMBuilderRef builder = bld->gallivm->builder;
251    LLVMValueRef res;
252 
253    assert(stencil[0].enabled);
254 
255    /* do front face op */
256    res = lp_build_stencil_op_single(bld, &stencil[0], op,
257                                      stencilRefs[0], stencilVals);
258 
259    if (stencil[1].enabled && front_facing != NULL) {
260       /* do back face op */
261       LLVMValueRef back_res;
262 
263       back_res = lp_build_stencil_op_single(bld, &stencil[1], op,
264                                             stencilRefs[1], stencilVals);
265 
266       res = lp_build_select(bld, front_facing, res, back_res);
267    }
268 
269    if (stencil[0].writemask != 0xff ||
270        (stencil[1].enabled && front_facing != NULL &&
271         stencil[1].writemask != 0xff)) {
272       /* mask &= stencil[0].writemask */
273       LLVMValueRef writemask = lp_build_const_int_vec(bld->gallivm, bld->type,
274                                                       stencil[0].writemask);
275       if (stencil[1].enabled &&
276           stencil[1].writemask != stencil[0].writemask &&
277           front_facing != NULL) {
278          LLVMValueRef back_writemask =
279             lp_build_const_int_vec(bld->gallivm, bld->type,
280                                    stencil[1].writemask);
281          writemask = lp_build_select(bld, front_facing,
282                                      writemask, back_writemask);
283       }
284 
285       mask = LLVMBuildAnd(builder, mask, writemask, "");
286       /* res = (res & mask) | (stencilVals & ~mask) */
287       res = lp_build_select_bitwise(bld, mask, res, stencilVals);
288    } else {
289       /* res = mask ? res : stencilVals */
290       res = lp_build_select(bld, mask, res, stencilVals);
291    }
292 
293    return res;
294 }
295 
296 
297 
298 /**
299  * Return a type that matches the depth/stencil format.
300  */
301 struct lp_type
lp_depth_type(const struct util_format_description * format_desc,unsigned length)302 lp_depth_type(const struct util_format_description *format_desc,
303               unsigned length)
304 {
305    struct lp_type type;
306    unsigned z_swizzle;
307 
308    assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS);
309    assert(format_desc->block.width == 1);
310    assert(format_desc->block.height == 1);
311 
312    memset(&type, 0, sizeof type);
313    type.width = format_desc->block.bits;
314 
315    z_swizzle = format_desc->swizzle[0];
316    if (z_swizzle < 4) {
317       if (format_desc->channel[z_swizzle].type == UTIL_FORMAT_TYPE_FLOAT) {
318          type.floating = true;
319          assert(z_swizzle == 0);
320          assert(format_desc->channel[z_swizzle].size == 32);
321       }
322       else if (format_desc->channel[z_swizzle].type == UTIL_FORMAT_TYPE_UNSIGNED) {
323          assert(format_desc->block.bits <= 32);
324          assert(format_desc->channel[z_swizzle].normalized);
325          if (format_desc->channel[z_swizzle].size < format_desc->block.bits) {
326             /* Prefer signed integers when possible, as SSE has less support
327              * for unsigned comparison;
328              */
329             type.sign = true;
330          }
331       }
332       else
333          assert(0);
334    }
335 
336    type.length = length;
337 
338    return type;
339 }
340 
341 
342 /**
343  * Compute bitmask and bit shift to apply to the incoming fragment Z values
344  * and the Z buffer values needed before doing the Z comparison.
345  *
346  * Note that we leave the Z bits in the position that we find them
347  * in the Z buffer (typically 0xffffff00 or 0x00ffffff).  That lets us
348  * get by with fewer bit twiddling steps.
349  */
350 static bool
get_z_shift_and_mask(const struct util_format_description * format_desc,unsigned * shift,unsigned * width,unsigned * mask)351 get_z_shift_and_mask(const struct util_format_description *format_desc,
352                      unsigned *shift, unsigned *width, unsigned *mask)
353 {
354    unsigned total_bits;
355    unsigned z_swizzle;
356 
357    assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS);
358    assert(format_desc->block.width == 1);
359    assert(format_desc->block.height == 1);
360 
361    /* 64bit d/s format is special already extracted 32 bits */
362    total_bits = format_desc->block.bits > 32 ? 32 : format_desc->block.bits;
363 
364    z_swizzle = format_desc->swizzle[0];
365 
366    if (z_swizzle == PIPE_SWIZZLE_NONE)
367       return false;
368 
369    *width = format_desc->channel[z_swizzle].size;
370    /* & 31 is for the same reason as the 32-bit limit above */
371    *shift = format_desc->channel[z_swizzle].shift & 31;
372 
373    if (*width == total_bits) {
374       *mask = 0xffffffff;
375    } else {
376       *mask = ((1 << *width) - 1) << *shift;
377    }
378 
379    return true;
380 }
381 
382 
383 /**
384  * Compute bitmask and bit shift to apply to the framebuffer pixel values
385  * to put the stencil bits in the least significant position.
386  * (i.e. 0x000000ff)
387  */
388 static bool
get_s_shift_and_mask(const struct util_format_description * format_desc,unsigned * shift,unsigned * mask)389 get_s_shift_and_mask(const struct util_format_description *format_desc,
390                      unsigned *shift, unsigned *mask)
391 {
392    const unsigned s_swizzle = format_desc->swizzle[1];
393 
394    if (s_swizzle == PIPE_SWIZZLE_NONE)
395       return false;
396 
397    /* just special case 64bit d/s format */
398    if (format_desc->block.bits > 32) {
399       /* XXX big-endian? */
400       assert(format_desc->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT);
401       *shift = 0;
402       *mask = 0xff;
403       return true;
404    }
405 
406    *shift = format_desc->channel[s_swizzle].shift;
407    const unsigned sz = format_desc->channel[s_swizzle].size;
408    *mask = (1U << sz) - 1U;
409 
410    return true;
411 }
412 
413 
414 /**
415  * Perform the occlusion test and increase the counter.
416  * Test the depth mask. Add the number of channel which has none zero mask
417  * into the occlusion counter. e.g. maskvalue is {-1, -1, -1, -1}.
418  * The counter will add 4.
419  * TODO: could get that out of the fs loop.
420  *
421  * \param type holds element type of the mask vector.
422  * \param maskvalue is the depth test mask.
423  * \param counter is a pointer of the uint32 counter.
424  */
425 void
lp_build_occlusion_count(struct gallivm_state * gallivm,struct lp_type type,LLVMValueRef maskvalue,LLVMValueRef counter)426 lp_build_occlusion_count(struct gallivm_state *gallivm,
427                          struct lp_type type,
428                          LLVMValueRef maskvalue,
429                          LLVMValueRef counter)
430 {
431    LLVMBuilderRef builder = gallivm->builder;
432    LLVMContextRef context = gallivm->context;
433    LLVMValueRef countmask = lp_build_const_int_vec(gallivm, type, 1);
434    LLVMValueRef count, newcount;
435 
436    assert(type.length <= 16);
437    assert(type.floating);
438 
439    if (util_get_cpu_caps()->has_sse && type.length == 4) {
440       const char *movmskintr = "llvm.x86.sse.movmsk.ps";
441       const char *popcntintr = "llvm.ctpop.i32";
442       LLVMValueRef bits = LLVMBuildBitCast(builder, maskvalue,
443                                            lp_build_vec_type(gallivm, type), "");
444       bits = lp_build_intrinsic_unary(builder, movmskintr,
445                                       LLVMInt32TypeInContext(context), bits);
446       count = lp_build_intrinsic_unary(builder, popcntintr,
447                                        LLVMInt32TypeInContext(context), bits);
448       count = LLVMBuildZExt(builder, count, LLVMIntTypeInContext(context, 64), "");
449    }
450    else if (util_get_cpu_caps()->has_avx && type.length == 8) {
451       const char *movmskintr = "llvm.x86.avx.movmsk.ps.256";
452       const char *popcntintr = "llvm.ctpop.i32";
453       LLVMValueRef bits = LLVMBuildBitCast(builder, maskvalue,
454                                            lp_build_vec_type(gallivm, type), "");
455       bits = lp_build_intrinsic_unary(builder, movmskintr,
456                                       LLVMInt32TypeInContext(context), bits);
457       count = lp_build_intrinsic_unary(builder, popcntintr,
458                                        LLVMInt32TypeInContext(context), bits);
459       count = LLVMBuildZExt(builder, count, LLVMIntTypeInContext(context, 64), "");
460    } else {
461       LLVMValueRef countv = LLVMBuildAnd(builder, maskvalue, countmask, "countv");
462       LLVMTypeRef counttype = LLVMIntTypeInContext(context, type.length * 8);
463       LLVMTypeRef i8vntype = LLVMVectorType(LLVMInt8TypeInContext(context), type.length * 4);
464       LLVMValueRef shufflev, countd;
465       LLVMValueRef shuffles[16];
466       const char *popcntintr = NULL;
467 
468       countv = LLVMBuildBitCast(builder, countv, i8vntype, "");
469 
470        for (unsigned i = 0; i < type.length; i++) {
471 #if UTIL_ARCH_LITTLE_ENDIAN
472           shuffles[i] = lp_build_const_int32(gallivm, 4*i);
473 #else
474           shuffles[i] = lp_build_const_int32(gallivm, (4*i) + 3);
475 #endif
476        }
477 
478        shufflev = LLVMConstVector(shuffles, type.length);
479        countd = LLVMBuildShuffleVector(builder, countv, LLVMGetUndef(i8vntype), shufflev, "");
480        countd = LLVMBuildBitCast(builder, countd, counttype, "countd");
481 
482        /*
483         * XXX FIXME
484         * this is bad on cpus without popcount (on x86 supported by intel
485         * nehalem, amd barcelona, and up - not tied to sse42).
486         * Would be much faster to just sum the 4 elements of the vector with
487         * some horizontal add (shuffle/add/shuffle/add after the initial and).
488         */
489        switch (type.length) {
490        case 4:
491           popcntintr = "llvm.ctpop.i32";
492           break;
493        case 8:
494           popcntintr = "llvm.ctpop.i64";
495           break;
496        case 16:
497           popcntintr = "llvm.ctpop.i128";
498           break;
499        default:
500           assert(0);
501        }
502        count = lp_build_intrinsic_unary(builder, popcntintr, counttype, countd);
503 
504        if (type.length > 8) {
505           count = LLVMBuildTrunc(builder, count, LLVMIntTypeInContext(context, 64), "");
506        }
507        else if (type.length < 8) {
508           count = LLVMBuildZExt(builder, count, LLVMIntTypeInContext(context, 64), "");
509        }
510    }
511    newcount = LLVMBuildLoad2(builder, LLVMTypeOf(count), counter, "origcount");
512    newcount = LLVMBuildAdd(builder, newcount, count, "newcount");
513    LLVMBuildStore(builder, newcount, counter);
514 }
515 
516 
517 /**
518  * Load depth/stencil values.
519  * The stored values are linear, swizzle them.
520  *
521  * \param type  the data type of the fragment depth/stencil values
522  * \param format_desc  description of the depth/stencil surface
523  * \param is_1d  whether this resource has only one dimension
524  * \param loop_counter  the current loop iteration
525  * \param depth_ptr  pointer to the depth/stencil values of this 4x4 block
526  * \param depth_stride  stride of the depth/stencil buffer
527  * \param z_fb  contains z values loaded from fb (may include padding)
528  * \param s_fb  contains s values loaded from fb (may include padding)
529  */
530 void
lp_build_depth_stencil_load_swizzled(struct gallivm_state * gallivm,struct lp_type z_src_type,const struct util_format_description * format_desc,bool is_1d,LLVMValueRef depth_ptr,LLVMValueRef depth_stride,LLVMValueRef * z_fb,LLVMValueRef * s_fb,LLVMValueRef loop_counter)531 lp_build_depth_stencil_load_swizzled(struct gallivm_state *gallivm,
532                                      struct lp_type z_src_type,
533                                      const struct util_format_description *format_desc,
534                                      bool is_1d,
535                                      LLVMValueRef depth_ptr,
536                                      LLVMValueRef depth_stride,
537                                      LLVMValueRef *z_fb,
538                                      LLVMValueRef *s_fb,
539                                      LLVMValueRef loop_counter)
540 {
541    LLVMBuilderRef builder = gallivm->builder;
542    LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 4];
543    LLVMValueRef depth_offset1, depth_offset2;
544    const unsigned depth_bytes = format_desc->block.bits / 8;
545    struct lp_type zs_type = lp_depth_type(format_desc, z_src_type.length);
546 
547    struct lp_type zs_load_type = zs_type;
548    zs_load_type.length = zs_load_type.length / 2;
549 
550    LLVMTypeRef zs_dst_type = lp_build_vec_type(gallivm, zs_load_type);
551 
552    if (z_src_type.length == 4) {
553       LLVMValueRef looplsb = LLVMBuildAnd(builder, loop_counter,
554                                           lp_build_const_int32(gallivm, 1), "");
555       LLVMValueRef loopmsb = LLVMBuildAnd(builder, loop_counter,
556                                           lp_build_const_int32(gallivm, 2), "");
557       LLVMValueRef offset2 = LLVMBuildMul(builder, loopmsb,
558                                           depth_stride, "");
559       depth_offset1 = LLVMBuildMul(builder, looplsb,
560                                    lp_build_const_int32(gallivm, depth_bytes * 2), "");
561       depth_offset1 = LLVMBuildAdd(builder, depth_offset1, offset2, "");
562 
563       /* just concatenate the loaded 2x2 values into 4-wide vector */
564       for (unsigned i = 0; i < 4; i++) {
565          shuffles[i] = lp_build_const_int32(gallivm, i);
566       }
567    } else {
568       unsigned i;
569       LLVMValueRef loopx2 = LLVMBuildShl(builder, loop_counter,
570                                          lp_build_const_int32(gallivm, 1), "");
571       assert(z_src_type.length == 8);
572       depth_offset1 = LLVMBuildMul(builder, loopx2, depth_stride, "");
573       /*
574        * We load 2x4 values, and need to swizzle them (order
575        * 0,1,4,5,2,3,6,7) - not so hot with avx unfortunately.
576        */
577       for (i = 0; i < 8; i++) {
578          shuffles[i] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2);
579       }
580    }
581 
582    depth_offset2 = LLVMBuildAdd(builder, depth_offset1, depth_stride, "");
583 
584    /* Load current z/stencil values from z/stencil buffer */
585    LLVMTypeRef load_ptr_type = LLVMPointerType(zs_dst_type, 0);
586    LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
587    LLVMValueRef zs_dst_ptr =
588       LLVMBuildGEP2(builder, int8_type, depth_ptr, &depth_offset1, 1, "");
589    zs_dst_ptr = LLVMBuildBitCast(builder, zs_dst_ptr, load_ptr_type, "");
590    LLVMValueRef zs_dst1 = LLVMBuildLoad2(builder, zs_dst_type, zs_dst_ptr, "");
591    LLVMValueRef zs_dst2;
592    if (is_1d) {
593       zs_dst2 = lp_build_undef(gallivm, zs_load_type);
594    } else {
595       zs_dst_ptr = LLVMBuildGEP2(builder, int8_type, depth_ptr, &depth_offset2, 1, "");
596       zs_dst_ptr = LLVMBuildBitCast(builder, zs_dst_ptr, load_ptr_type, "");
597       zs_dst2 = LLVMBuildLoad2(builder, zs_dst_type, zs_dst_ptr, "");
598    }
599 
600    *z_fb = LLVMBuildShuffleVector(builder, zs_dst1, zs_dst2,
601                                   LLVMConstVector(shuffles, zs_type.length), "");
602    *s_fb = *z_fb;
603 
604    if (format_desc->block.bits == 8) {
605       /* Extend stencil-only 8 bit values (S8_UINT) */
606       *s_fb = LLVMBuildZExt(builder, *s_fb,
607                             lp_build_int_vec_type(gallivm, z_src_type), "");
608    }
609 
610    if (format_desc->block.bits < z_src_type.width) {
611       /* Extend destination ZS values (e.g., when reading from Z16_UNORM) */
612       *z_fb = LLVMBuildZExt(builder, *z_fb,
613                             lp_build_int_vec_type(gallivm, z_src_type), "");
614    }
615 
616    else if (format_desc->block.bits > 32) {
617       /* rely on llvm to handle too wide vector we have here nicely */
618       struct lp_type typex2 = zs_type;
619       struct lp_type s_type = zs_type;
620       LLVMValueRef shuffles1[LP_MAX_VECTOR_LENGTH / 4];
621       LLVMValueRef shuffles2[LP_MAX_VECTOR_LENGTH / 4];
622       LLVMValueRef tmp;
623 
624       typex2.width = typex2.width / 2;
625       typex2.length = typex2.length * 2;
626       s_type.width = s_type.width / 2;
627       s_type.floating = 0;
628 
629       tmp = LLVMBuildBitCast(builder, *z_fb,
630                              lp_build_vec_type(gallivm, typex2), "");
631 
632       for (unsigned i = 0; i < zs_type.length; i++) {
633          shuffles1[i] = lp_build_const_int32(gallivm, i * 2);
634          shuffles2[i] = lp_build_const_int32(gallivm, i * 2 + 1);
635       }
636       *z_fb = LLVMBuildShuffleVector(builder, tmp, tmp,
637                                      LLVMConstVector(shuffles1, zs_type.length), "");
638       *s_fb = LLVMBuildShuffleVector(builder, tmp, tmp,
639                                      LLVMConstVector(shuffles2, zs_type.length), "");
640       *s_fb = LLVMBuildBitCast(builder, *s_fb,
641                                lp_build_vec_type(gallivm, s_type), "");
642       lp_build_name(*s_fb, "s_dst");
643    }
644 
645    lp_build_name(*z_fb, "z_dst");
646    lp_build_name(*s_fb, "s_dst");
647    lp_build_name(*z_fb, "z_dst");
648 }
649 
650 
651 /**
652  * Store depth/stencil values.
653  * Incoming values are swizzled (typically n 2x2 quads), stored linear.
654  * If there's a mask it will do select/store otherwise just store.
655  *
656  * \param type  the data type of the fragment depth/stencil values
657  * \param format_desc  description of the depth/stencil surface
658  * \param is_1d  whether this resource has only one dimension
659  * \param mask_value the alive/dead pixel mask for the quad (vector)
660  * \param z_fb  z values read from fb (with padding)
661  * \param s_fb  s values read from fb (with padding)
662  * \param loop_counter  the current loop iteration
663  * \param depth_ptr  pointer to the depth/stencil values of this 4x4 block
664  * \param depth_stride  stride of the depth/stencil buffer
665  * \param z_value the depth values to store (with padding)
666  * \param s_value the stencil values to store (with padding)
667  */
668 void
lp_build_depth_stencil_write_swizzled(struct gallivm_state * gallivm,struct lp_type z_src_type,const struct util_format_description * format_desc,bool is_1d,LLVMValueRef mask_value,LLVMValueRef z_fb,LLVMValueRef s_fb,LLVMValueRef loop_counter,LLVMValueRef depth_ptr,LLVMValueRef depth_stride,LLVMValueRef z_value,LLVMValueRef s_value)669 lp_build_depth_stencil_write_swizzled(struct gallivm_state *gallivm,
670                                       struct lp_type z_src_type,
671                                       const struct util_format_description *format_desc,
672                                       bool is_1d,
673                                       LLVMValueRef mask_value,
674                                       LLVMValueRef z_fb,
675                                       LLVMValueRef s_fb,
676                                       LLVMValueRef loop_counter,
677                                       LLVMValueRef depth_ptr,
678                                       LLVMValueRef depth_stride,
679                                       LLVMValueRef z_value,
680                                       LLVMValueRef s_value)
681 {
682    struct lp_build_context z_bld;
683    LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 4];
684    LLVMBuilderRef builder = gallivm->builder;
685    LLVMValueRef zs_dst1, zs_dst2;
686    LLVMValueRef zs_dst_ptr1, zs_dst_ptr2;
687    LLVMValueRef depth_offset1, depth_offset2;
688    LLVMTypeRef load_ptr_type;
689    unsigned depth_bytes = format_desc->block.bits / 8;
690    struct lp_type zs_type = lp_depth_type(format_desc, z_src_type.length);
691    struct lp_type z_type = zs_type;
692    struct lp_type zs_load_type = zs_type;
693 
694    zs_load_type.length = zs_load_type.length / 2;
695    load_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, zs_load_type), 0);
696 
697    z_type.width = z_src_type.width;
698 
699    lp_build_context_init(&z_bld, gallivm, z_type);
700 
701    /*
702     * This is far from ideal, at least for late depth write we should do this
703     * outside the fs loop to avoid all the swizzle stuff.
704     */
705    if (z_src_type.length == 4) {
706       LLVMValueRef looplsb = LLVMBuildAnd(builder, loop_counter,
707                                           lp_build_const_int32(gallivm, 1), "");
708       LLVMValueRef loopmsb = LLVMBuildAnd(builder, loop_counter,
709                                           lp_build_const_int32(gallivm, 2), "");
710       LLVMValueRef offset2 = LLVMBuildMul(builder, loopmsb,
711                                           depth_stride, "");
712       depth_offset1 = LLVMBuildMul(builder, looplsb,
713                                    lp_build_const_int32(gallivm, depth_bytes * 2), "");
714       depth_offset1 = LLVMBuildAdd(builder, depth_offset1, offset2, "");
715    } else {
716       LLVMValueRef loopx2 = LLVMBuildShl(builder, loop_counter,
717                                          lp_build_const_int32(gallivm, 1), "");
718       assert(z_src_type.length == 8);
719       depth_offset1 = LLVMBuildMul(builder, loopx2, depth_stride, "");
720       /*
721        * We load 2x4 values, and need to swizzle them (order
722        * 0,1,4,5,2,3,6,7) - not so hot with avx unfortunately.
723        */
724       for (unsigned i = 0; i < 8; i++) {
725          shuffles[i] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2);
726       }
727    }
728 
729    depth_offset2 = LLVMBuildAdd(builder, depth_offset1, depth_stride, "");
730 
731    LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
732    zs_dst_ptr1 = LLVMBuildGEP2(builder, int8_type, depth_ptr, &depth_offset1, 1, "");
733    zs_dst_ptr1 = LLVMBuildBitCast(builder, zs_dst_ptr1, load_ptr_type, "");
734    zs_dst_ptr2 = LLVMBuildGEP2(builder, int8_type, depth_ptr, &depth_offset2, 1, "");
735    zs_dst_ptr2 = LLVMBuildBitCast(builder, zs_dst_ptr2, load_ptr_type, "");
736 
737    if (format_desc->block.bits > 32) {
738       s_value = LLVMBuildBitCast(builder, s_value, z_bld.vec_type, "");
739    }
740 
741    if (mask_value) {
742       z_value = lp_build_select(&z_bld, mask_value, z_value, z_fb);
743       if (format_desc->block.bits > 32) {
744          s_fb = LLVMBuildBitCast(builder, s_fb, z_bld.vec_type, "");
745          s_value = lp_build_select(&z_bld, mask_value, s_value, s_fb);
746       }
747    }
748 
749    if (zs_type.width < z_src_type.width) {
750       /* Truncate ZS values (e.g., when writing to Z16_UNORM) */
751       z_value = LLVMBuildTrunc(builder, z_value,
752                                lp_build_int_vec_type(gallivm, zs_type), "");
753    }
754 
755    if (format_desc->block.bits <= 32) {
756       if (z_src_type.length == 4) {
757          zs_dst1 = lp_build_extract_range(gallivm, z_value, 0, 2);
758          zs_dst2 = lp_build_extract_range(gallivm, z_value, 2, 2);
759       } else {
760          assert(z_src_type.length == 8);
761          zs_dst1 = LLVMBuildShuffleVector(builder, z_value, z_value,
762                                           LLVMConstVector(&shuffles[0],
763                                                           zs_load_type.length), "");
764          zs_dst2 = LLVMBuildShuffleVector(builder, z_value, z_value,
765                                           LLVMConstVector(&shuffles[4],
766                                                           zs_load_type.length), "");
767       }
768    } else {
769       if (z_src_type.length == 4) {
770          zs_dst1 = lp_build_interleave2(gallivm, z_type,
771                                         z_value, s_value, 0);
772          zs_dst2 = lp_build_interleave2(gallivm, z_type,
773                                         z_value, s_value, 1);
774       } else {
775          LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 2];
776          assert(z_src_type.length == 8);
777          for (unsigned i = 0; i < 8; i++) {
778             shuffles[i*2] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2);
779             shuffles[i*2+1] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2 +
780                                                    z_src_type.length);
781          }
782          zs_dst1 = LLVMBuildShuffleVector(builder, z_value, s_value,
783                                           LLVMConstVector(&shuffles[0],
784                                                           z_src_type.length), "");
785          zs_dst2 = LLVMBuildShuffleVector(builder, z_value, s_value,
786                                           LLVMConstVector(&shuffles[8],
787                                                           z_src_type.length), "");
788       }
789       zs_dst1 = LLVMBuildBitCast(builder, zs_dst1,
790                                  lp_build_vec_type(gallivm, zs_load_type), "");
791       zs_dst2 = LLVMBuildBitCast(builder, zs_dst2,
792                                  lp_build_vec_type(gallivm, zs_load_type), "");
793    }
794 
795    LLVMBuildStore(builder, zs_dst1, zs_dst_ptr1);
796    if (!is_1d) {
797       LLVMBuildStore(builder, zs_dst2, zs_dst_ptr2);
798    }
799 }
800 
801 
802 /**
803  * Generate code for performing depth and/or stencil tests.
804  * We operate on a vector of values (typically n 2x2 quads).
805  *
806  * \param depth  the depth test state
807  * \param stencil  the front/back stencil state
808  * \param type  the data type of the fragment depth/stencil values
809  * \param format_desc  description of the depth/stencil surface
810  * \param mask  the alive/dead pixel mask for the quad (vector)
811  * \param cov_mask coverage mask
812  * \param stencil_refs  the front/back stencil ref values (scalar)
813  * \param z_src  the incoming depth/stencil values (n 2x2 quad values, float32)
814  * \param zs_dst  the depth/stencil values in framebuffer
815  * \param face  contains boolean value indicating front/back facing polygon
816  */
817 void
lp_build_depth_stencil_test(struct gallivm_state * gallivm,const struct lp_depth_state * depth,const struct pipe_stencil_state stencil[2],struct lp_type z_src_type,const struct util_format_description * format_desc,struct lp_build_mask_context * mask,LLVMValueRef * cov_mask,LLVMValueRef stencil_refs[2],LLVMValueRef z_src,LLVMValueRef z_fb,LLVMValueRef s_fb,LLVMValueRef face,LLVMValueRef * z_value,LLVMValueRef * s_value,bool do_branch,bool restrict_depth)818 lp_build_depth_stencil_test(struct gallivm_state *gallivm,
819                             const struct lp_depth_state *depth,
820                             const struct pipe_stencil_state stencil[2],
821                             struct lp_type z_src_type,
822                             const struct util_format_description *format_desc,
823                             struct lp_build_mask_context *mask,
824                             LLVMValueRef *cov_mask,
825                             LLVMValueRef stencil_refs[2],
826                             LLVMValueRef z_src,
827                             LLVMValueRef z_fb,
828                             LLVMValueRef s_fb,
829                             LLVMValueRef face,
830                             LLVMValueRef *z_value,
831                             LLVMValueRef *s_value,
832                             bool do_branch,
833                             bool restrict_depth)
834 {
835    LLVMBuilderRef builder = gallivm->builder;
836    struct lp_type z_type;
837    struct lp_build_context z_bld;
838    struct lp_build_context s_bld;
839    struct lp_type s_type;
840    unsigned z_shift = 0, z_width = 0, z_mask = 0;
841    LLVMValueRef z_dst = NULL;
842    LLVMValueRef stencil_vals = NULL;
843    LLVMValueRef z_bitmask = NULL, stencil_shift = NULL;
844    LLVMValueRef z_pass = NULL, s_pass_mask = NULL;
845    LLVMValueRef current_mask = mask ? lp_build_mask_value(mask) : *cov_mask;
846    LLVMValueRef front_facing = NULL;
847    bool have_z, have_s;
848 
849    /*
850     * Depths are expected to be between 0 and 1, even if they are stored in
851     * floats. Setting these bits here will ensure that the lp_build_conv() call
852     * below won't try to unnecessarily clamp the incoming values.
853     * If depths are expected outside 0..1 don't set these bits.
854     */
855    if (z_src_type.floating) {
856       if (restrict_depth) {
857          z_src_type.sign = false;
858          z_src_type.norm = true;
859       }
860    } else {
861       assert(!z_src_type.sign);
862       assert(z_src_type.norm);
863    }
864 
865    /* Pick the type matching the depth-stencil format. */
866    z_type = lp_depth_type(format_desc, z_src_type.length);
867 
868    /* Pick the intermediate type for depth operations. */
869    z_type.width = z_src_type.width;
870    assert(z_type.length == z_src_type.length);
871 
872    /* FIXME: for non-float depth/stencil might generate better code
873     * if we'd always split it up to use 128bit operations.
874     * For stencil we'd almost certainly want to pack to 8xi16 values,
875     * for z just run twice.
876     */
877 
878    /* Sanity checking */
879    {
880       ASSERTED const unsigned z_swizzle = format_desc->swizzle[0];
881       ASSERTED const unsigned s_swizzle = format_desc->swizzle[1];
882 
883       assert(z_swizzle != PIPE_SWIZZLE_NONE ||
884              s_swizzle != PIPE_SWIZZLE_NONE);
885 
886       assert(depth->enabled || stencil[0].enabled);
887 
888       assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS);
889       assert(format_desc->block.width == 1);
890       assert(format_desc->block.height == 1);
891 
892       if (stencil[0].enabled) {
893          assert(s_swizzle < 4);
894          assert(format_desc->channel[s_swizzle].type == UTIL_FORMAT_TYPE_UNSIGNED);
895          assert(format_desc->channel[s_swizzle].pure_integer);
896          assert(!format_desc->channel[s_swizzle].normalized);
897          assert(format_desc->channel[s_swizzle].size == 8);
898       }
899 
900       if (depth->enabled) {
901          assert(z_swizzle < 4);
902          if (z_type.floating) {
903             assert(z_swizzle == 0);
904             assert(format_desc->channel[z_swizzle].type ==
905                    UTIL_FORMAT_TYPE_FLOAT);
906             assert(format_desc->channel[z_swizzle].size == 32);
907          } else {
908             assert(format_desc->channel[z_swizzle].type ==
909                    UTIL_FORMAT_TYPE_UNSIGNED);
910             assert(format_desc->channel[z_swizzle].normalized);
911             assert(!z_type.fixed);
912          }
913       }
914    }
915 
916 
917    /* Setup build context for Z vals */
918    lp_build_context_init(&z_bld, gallivm, z_type);
919 
920    /* Setup build context for stencil vals */
921    s_type = lp_int_type(z_type);
922    lp_build_context_init(&s_bld, gallivm, s_type);
923 
924    /* Compute and apply the Z/stencil bitmasks and shifts.
925     */
926    {
927       unsigned s_shift, s_mask;
928 
929       z_dst = z_fb;
930       stencil_vals = s_fb;
931 
932       have_z = get_z_shift_and_mask(format_desc, &z_shift, &z_width, &z_mask);
933       have_s = get_s_shift_and_mask(format_desc, &s_shift, &s_mask);
934 
935       if (have_z) {
936          if (z_mask != 0xffffffff) {
937             z_bitmask = lp_build_const_int_vec(gallivm, z_type, z_mask);
938          }
939 
940          /*
941           * Align the framebuffer Z 's LSB to the right.
942           */
943          if (z_shift) {
944             LLVMValueRef shift = lp_build_const_int_vec(gallivm, z_type, z_shift);
945             z_dst = LLVMBuildLShr(builder, z_dst, shift, "z_dst");
946          } else if (z_bitmask) {
947             z_dst = LLVMBuildAnd(builder, z_dst, z_bitmask, "z_dst");
948          } else {
949             lp_build_name(z_dst, "z_dst");
950          }
951       }
952 
953       if (have_s) {
954          if (s_shift) {
955             LLVMValueRef shift = lp_build_const_int_vec(gallivm, s_type, s_shift);
956             stencil_vals = LLVMBuildLShr(builder, stencil_vals, shift, "");
957             stencil_shift = shift;  /* used below */
958          }
959 
960          if (s_mask != 0xffffffff) {
961             LLVMValueRef mask = lp_build_const_int_vec(gallivm, s_type, s_mask);
962             stencil_vals = LLVMBuildAnd(builder, stencil_vals, mask, "");
963          }
964 
965          lp_build_name(stencil_vals, "s_dst");
966       }
967    }
968 
969    if (stencil[0].enabled) {
970 
971       if (face) {
972          if (0) {
973             /*
974              * XXX: the scalar expansion below produces atrocious code
975              * (basically producing a 64bit scalar value, then moving the 2
976              * 32bit pieces separately to simd, plus 4 shuffles, which is
977              * seriously lame). But the scalar-simd transitions are always
978              * tricky, so no big surprise there.
979              * This here would be way better, however llvm has some serious
980              * trouble later using it in the select, probably because it will
981              * recognize the expression as constant and move the simd value
982              * away (out of the loop) - and then it will suddenly try
983              * constructing i1 high-bit masks out of it later...
984              * (Try piglit stencil-twoside.)
985              * Note this is NOT due to using SExt/Trunc, it fails exactly the
986              * same even when using native compare/select.
987              * I cannot reproduce this problem when using stand-alone compiler
988              * though, suggesting some problem with optimization passes...
989              * (With stand-alone compilation, the construction of this mask
990              * value, no matter if the easy 3 instruction here or the complex
991              * 16+ one below, never gets separated from where it's used.)
992              * The scalar code still has the same problem, but the generated
993              * code looks a bit better at least for some reason, even if
994              * mostly by luck (the fundamental issue clearly is the same).
995              */
996             front_facing = lp_build_broadcast(gallivm, s_bld.vec_type, face);
997             /* front_facing = face != 0 ? ~0 : 0 */
998             front_facing = lp_build_compare(gallivm, s_bld.type,
999                                             PIPE_FUNC_NOTEQUAL,
1000                                             front_facing, s_bld.zero);
1001          } else {
1002             LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1003 
1004             /* front_facing = face != 0 ? ~0 : 0 */
1005             front_facing = LLVMBuildICmp(builder, LLVMIntNE, face, zero, "");
1006             front_facing = LLVMBuildSExt(builder, front_facing,
1007                                          LLVMIntTypeInContext(gallivm->context,
1008                                                 s_bld.type.length*s_bld.type.width),
1009                                          "");
1010             front_facing = LLVMBuildBitCast(builder, front_facing,
1011                                             s_bld.int_vec_type, "");
1012 
1013          }
1014       }
1015 
1016       s_pass_mask = lp_build_stencil_test(&s_bld, stencil,
1017                                           stencil_refs, stencil_vals,
1018                                           front_facing);
1019 
1020       /* apply stencil-fail operator */
1021       {
1022          LLVMValueRef s_fail_mask = lp_build_andnot(&s_bld, current_mask, s_pass_mask);
1023          stencil_vals = lp_build_stencil_op(&s_bld, stencil, S_FAIL_OP,
1024                                             stencil_refs, stencil_vals,
1025                                             s_fail_mask, front_facing);
1026       }
1027    }
1028 
1029    if (depth->enabled) {
1030       /*
1031        * Convert fragment Z to the desired type, aligning the LSB to the right.
1032        */
1033 
1034       assert(z_type.width == z_src_type.width);
1035       assert(z_type.length == z_src_type.length);
1036       assert(lp_check_value(z_src_type, z_src));
1037       if (z_src_type.floating) {
1038          /*
1039           * Convert from floating point values
1040           */
1041 
1042          if (!z_type.floating) {
1043             z_src = lp_build_clamped_float_to_unsigned_norm(gallivm,
1044                                                             z_src_type,
1045                                                             z_width,
1046                                                             z_src);
1047          }
1048       } else {
1049          /*
1050           * Convert from unsigned normalized values.
1051           */
1052 
1053          assert(!z_src_type.sign);
1054          assert(!z_src_type.fixed);
1055          assert(z_src_type.norm);
1056          assert(!z_type.floating);
1057          if (z_src_type.width > z_width) {
1058             LLVMValueRef shift = lp_build_const_int_vec(gallivm, z_src_type,
1059                                                         z_src_type.width - z_width);
1060             z_src = LLVMBuildLShr(builder, z_src, shift, "");
1061          }
1062       }
1063       assert(lp_check_value(z_type, z_src));
1064 
1065       lp_build_name(z_src, "z_src");
1066 
1067       /* compare src Z to dst Z, returning 'pass' mask */
1068       z_pass = lp_build_cmp(&z_bld, depth->func, z_src, z_dst);
1069 
1070       /* mask off bits that failed stencil test */
1071       if (s_pass_mask) {
1072          current_mask = LLVMBuildAnd(builder, current_mask, s_pass_mask, "");
1073       }
1074 
1075       if (!stencil[0].enabled && mask) {
1076          /* We can potentially skip all remaining operations here, but only
1077           * if stencil is disabled because we still need to update the stencil
1078           * buffer values.  Don't need to update Z buffer values.
1079           */
1080          lp_build_mask_update(mask, z_pass);
1081 
1082          if (do_branch) {
1083             lp_build_mask_check(mask);
1084          }
1085       }
1086 
1087       if (depth->writemask) {
1088          LLVMValueRef z_pass_mask;
1089 
1090          /* mask off bits that failed Z test */
1091          z_pass_mask = LLVMBuildAnd(builder, current_mask, z_pass, "");
1092 
1093          /* Mix the old and new Z buffer values.
1094           * z_dst[i] = zselectmask[i] ? z_src[i] : z_dst[i]
1095           */
1096          z_dst = lp_build_select(&z_bld, z_pass_mask, z_src, z_dst);
1097       }
1098 
1099       if (stencil[0].enabled) {
1100          /* update stencil buffer values according to z pass/fail result */
1101          LLVMValueRef z_fail_mask, z_pass_mask;
1102 
1103          /* apply Z-fail operator */
1104          z_fail_mask = lp_build_andnot(&s_bld, current_mask, z_pass);
1105          stencil_vals = lp_build_stencil_op(&s_bld, stencil, Z_FAIL_OP,
1106                                             stencil_refs, stencil_vals,
1107                                             z_fail_mask, front_facing);
1108 
1109          /* apply Z-pass operator */
1110          z_pass_mask = LLVMBuildAnd(builder, current_mask, z_pass, "");
1111          stencil_vals = lp_build_stencil_op(&s_bld, stencil, Z_PASS_OP,
1112                                             stencil_refs, stencil_vals,
1113                                             z_pass_mask, front_facing);
1114       }
1115    } else {
1116       /* No depth test: apply Z-pass operator to stencil buffer values which
1117        * passed the stencil test.
1118        */
1119       s_pass_mask = LLVMBuildAnd(builder, current_mask, s_pass_mask, "");
1120       stencil_vals = lp_build_stencil_op(&s_bld, stencil, Z_PASS_OP,
1121                                          stencil_refs, stencil_vals,
1122                                          s_pass_mask, front_facing);
1123    }
1124 
1125    /* Put Z and stencil bits in the right place */
1126    if (have_z && z_shift) {
1127       LLVMValueRef shift = lp_build_const_int_vec(gallivm, z_type, z_shift);
1128       z_dst = LLVMBuildShl(builder, z_dst, shift, "");
1129    }
1130    if (stencil_vals && stencil_shift)
1131       stencil_vals = LLVMBuildShl(builder, stencil_vals,
1132                                   stencil_shift, "");
1133 
1134    /* Finally, merge the z/stencil values */
1135    if (format_desc->block.bits <= 32) {
1136       if (have_z && have_s)
1137          *z_value = LLVMBuildOr(builder, z_dst, stencil_vals, "");
1138       else if (have_z)
1139          *z_value = z_dst;
1140       else
1141          *z_value = stencil_vals;
1142       *s_value = *z_value;
1143    } else {
1144       *z_value = z_dst;
1145       *s_value = stencil_vals;
1146    }
1147 
1148    if (mask) {
1149       if (s_pass_mask)
1150          lp_build_mask_update(mask, s_pass_mask);
1151 
1152       if (depth->enabled && stencil[0].enabled)
1153          lp_build_mask_update(mask, z_pass);
1154    } else {
1155       LLVMValueRef tmp_mask = *cov_mask;
1156       if (s_pass_mask)
1157          tmp_mask = LLVMBuildAnd(builder, tmp_mask, s_pass_mask, "");
1158 
1159       /* for multisample we don't do the stencil optimisation so update always */
1160       if (depth->enabled)
1161          tmp_mask = LLVMBuildAnd(builder, tmp_mask, z_pass, "");
1162       *cov_mask = tmp_mask;
1163    }
1164 }
1165 
1166